Demucs.net is an independent, unofficial web interface for running the open-source Demucs music source separation model. It is designed for people who want to split a mixed song into parts without installing Python, PyTorch or CUDA on their own computer. The model runs through a hosted service: the visitor supplies an audio file, selects separation settings, starts a job and receives separate stems for playback and download. Demucs.net is a third-party runtime, not the Demucs project, and it is not affiliated with, endorsed by or sponsored by Meta or the Demucs authors.
Visitors can upload MP3, WAV, FLAC, M4A or OGG audio. The interface checks file length and format before upload. Depending on the selected model and mode, a job can produce vocals and accompaniment or additional stems such as drums, bass and other sounds. This supports use cases such as practicing an instrumental part, preparing a karaoke backing track, or studying how a recording is arranged. Stem playback is available in the browser, and individual tracks can be downloaded. Some paid options include lossless WAV and FLAC output; available formats and limits depend on the account plan.
An account is required before a separation starts. New accounts receive monthly free credits after Google sign-in or email confirmation, with no card needed for the free tier. Paid plans and credit packs offer more credits and capacity. The number of credits for a job depends on the audio length and the selected separation model and stem mode. A failed job returns held credits rather than charging for a result that was not delivered.
Demucs.net links to the upstream open-source repository and publishes a licensing explanation for the code and pretrained weights. It does not claim to be the official Demucs website, to improve the underlying model's separation quality, or to grant rights to use third-party recordings. Users must have appropriate rights to the audio they upload and to their use of the resulting stems. Separation quality depends on the source recording and model output, so no particular result is guaranteed.
